[Linux-cluster] 2nd try: fencing?

[Eric Kerin]

On Fri, 2006-04-28 at 21:58 -0400, Jason wrote:
> ok, so I have GFS installed and have the modules loaded, seems the next step is to set up "fencing"
> and looking at the list of agents at,
> http://www.redhat.com/docs/manuals/csgfs/admin-guide/s1-fence-methods.html
> im kinda confused. I thought that the idea behind fencing was to allow one node (i have 2) to take 
> down/reboot the other node after it has detected some sort of fault in the other node.

Quick and dirty reason for needing fencing: In GFS you need to stop a
failed (or semi-failed) node from writing data to the shared filesystem,
otherwise corruption may occur.  

One method is to power off a misbehaving node.  Another is to block
access to the shared disk by telling a SAN switch to disable it's port.
Still another is to tell a firewall to not allow network traffic to an
iSCSI device from the offending node.

What you use for a fence method all depends on your hardware.  If you
give a quick explanation of your hardware setup, we might be able to
help you pick a fence device that will work with what you have already.
Or if you don't have anything that could be used to block access, you
might have to buy some network power switches. 

OR, if this isn't intended for production use, and you're just testing,
you can use fence_manual.  This one has the unpleasant downside of
needing manual intervention to bring the cluster up after a node
failure.  But for testing GFS and Cluster Suite, it's nice and cheap.

Thanks, 
Eric Kerin
eric at bootseg.com